HOG特征向量的代码 源代码MATLAB

[html] view plain copy

1. <pre name="code" class="cpp">function HOGFeature = ImgHOGFeature( imgPathName, cell_size, nblock,...,  
2.     overlap, angle, bin_num)  
3. % 计算输入图像的特征描述子  
4. % imgPathName：图片路径  
5. % cell_size: cell的长宽  
6. % nblock: block的width、height包括的cell的个数  
7. % overlap: block重叠的比例  
8. % angle: 180\360  
9. % bin_num: 方向bin的数目  
10.    
11. if nargin<2  
12.     % default parameter  
13.     cell_size=8;  
14.     nblock=2;  
15.     overlap=0.5;  
16.     angle=180;  
17.     bin_num=9;  
18. elseif nargin<6  
19.     error('Input parameters are not enough.');  
20. end  
21.    
22. Img = imread(imgPathName);  
23. if size(Img,3) == 3  
24.     % 简化计算，直接转换成灰度图像  
25.     G = rgb2gray(Img);  
26. else  
27.     G = Img;  
28. end  
29.    
30. [height, width] = size(G);  
31.    
32. % 计算x、y方向的梯度  
33. hx = [-1,0,1];  
34. hy = -hx';  
35. grad_x = imfilter(double(G),hx);  
36. grad_y = imfilter(double(G),hy);  
37.    
38. % 计算梯度的模长  
39. grad_mag=sqrt(grad_x.^2+grad_y.^2);  
40.    
41. % 计算梯度的方向  
42. index= grad_x==0;  
43. grad_x(index)=1e-5;  
44. YX=grad_y./grad_x;  
45. if angle==180  
46.     grad_angle= ((atan(YX)+(pi/2))*180)./pi;   
47. elseif angle==360  
48.     grad_angle= ((atan2(grad_y,grad_x)+pi).*180)./pi;  
49. end  
50.    
51. % orient bin  
52. bin_angle=angle/bin_num;  
53. grad_orient=ceil(grad_angle./bin_angle);  
54.    
55. % 计算block的个数  
56. block_size=cell_size*nblock;  
57. skip_step=block_size*overlap;  
58. x_step_num=floor((width-block_size)/skip_step+1);  
59. y_step_num=floor((height-block_size)/skip_step+1);  
60.    
61. % 初始化hog特征描述子  
62. feat_dim=bin_num*nblock^2;  
63. HOGFeature=zeros(feat_dim,x_step_num*y_step_num);  
64.    
65. for k=1:y_step_num  
66.     for j=1:x_step_num  
67.         % block的左上角坐标  
68.         x_off = (j-1)*skip_step+1;  
69.         y_off = (k-1)*skip_step+1;  
70.    
71.         % 取得block的梯度大小和方向  
72.         b_mag=grad_mag(y_off:y_off+block_size-1,x_off:x_off+block_size-1);  
73.         b_orient=grad_orient(y_off:y_off+block_size-1,x_off:x_off+block_size-1);  
74.    
75.         % 当前block的hog直方图  
76.         currFeat = BinHOGFeature(b_mag, b_orient, cell_size,nblock, bin_num, false);  
77.         HOGFeature(:, (k-1)*x_step_num+j) = currFeat;  
78.    
79.     end  
80. end  
81.    
82. end  
83. function blockfeat = BinHOGFeature( b_mag,b_orient,cell_size,nblock,...  
84.     bin_num, weight_vote)  
85. % 计算1个block的hog  
86. % weight_vote: 是否进行高斯加权投票  
87.    
88. % block的HOG直方图  
89. blockfeat=zeros(bin_num*nblock^2,1);  
90.    
91. % 高斯权重  
92. gaussian_weight=fspecial('gaussian',cell_size*nblock,0.5*cell_size*nblock);  
93.    
94. % 分割block  
95. for n=1:nblock  
96.     for m=1:nblock  
97.         % cell的左上角坐标  
98.         x_off = (m-1)*cell_size+1;  
99.         y_off = (n-1)*cell_size+1;  
100.    
101.         % cell的梯度大小和方向  
102.         c_mag=b_mag(y_off:y_off+cell_size-1,x_off:x_off+cell_size-1);  
103.         c_orient=b_orient(y_off:y_off+cell_size-1,x_off:x_off+cell_size-1);  
104.    
105.         % cell的hog直方图  
106.         c_feat=zeros(bin_num,1);  
107.         for i=1:bin_num  
108.             % 是否进行高斯加权 投票  
109.             if weight_vote==false  
110.                 c_feat(i)=sum(c_mag(c_orient==i));  
111.             else  
112.                 c_feat(i)=sum(c_mag(c_orient==i).*gaussian_weight(c_orient==i));  
113.             end  
114.         end  
115.    
116.         % 合并到block的HOG直方图中  
117.         count=(n-1)*nblock+m;  
118.         blockfeat((count-1)*bin_num+1:count*bin_num,1)=c_feat;  
119.     end  
120. end  
121.    
122. % 归一化 L2-norm  
123. sump=sum(blockfeat.^2);  
124. blockfeat = blockfeat./sqrt(sump+eps^2);